Grok vs ChatGPT: cheaper answers, and a long wait

Pick ChatGPT (GPT-5.5) for anything interactive. Every GPT-5.5 answer arrived in under 9 seconds; no Grok answer took less than 33. GPT-5.5 also wrote the better email and the more careful code fix, and it named its assumptions where Grok just made them.

Pick Grok (Grok 4.3) if you run many requests in the background, through the API, and read the results later. At its list price of $1.25 in and $2.50 out per million tokens it is the cheapest of the five models we tested, and in our run its plan was the one that respected the budget. Its writing is short to the point of terse, which suits notes better than letters.

On half the prompts we could not separate them. Both explained compound interest correctly, both summarised the memo without inventing facts, and both answered the trick puzzle in two sentences.

We sent each prompt once to openai/gpt-5.5 and x-ai/grok-4.3 through OpenRouter's API on 7 November 2026, between 16:34 and 16:38 UTC. OpenRouter passed the Grok calls to xAI and the GPT calls to OpenAI. There was no system prompt, no web search and no chat history, and each answer was capped at 700 tokens. Both models think before they answer; we set both to the low reasoning setting.

The same prompt can come back differently tomorrow, so take one run as one data point. Where an answer was long we cut it ([...]) and printed its dashes as commas, and changed nothing else. None of these calls went through AI4Chat's own chat.

We asked for a 3-day Jaipur plan for two on ₹15,000, train tickets excluded, with a cost for each day. This is the prompt that tripped up most of the models we tried, because the totals have to agree.

Grok's days came to ₹4,800, ₹5,200 and ₹4,500, and its closing line broke the same money down by category: “Total: ₹14,500 (hotel ₹5,400 for 3 nights, food ₹4,500, transport ₹2,800, entries/misc ₹1,800).” Both ways it adds up to ₹14,500. It priced three hotel nights, so there is slack if you only need two.

GPT-5.5's days added up correctly too, to ₹17,000, which is ₹2,000 over. It then wrote “To keep it within ₹15,000, reduce shopping by ₹1,000 and choose shared autos/public transport to save around ₹1,000.” That is a note on a plan, not a plan.

Our call: Grok. We checked that each plan agrees with itself, not whether the hotel and entry prices are right.

The prompt was a blunt note to a client, Megan, about an $850 invoice three weeks late, to be rewritten "polite but firm" in under 120 words.

Grok's version was 42 words and stopped at the request: “I've sent two prior reminders without response. Could you please confirm when payment will be made?” It closed with "Thank you," and no name. There is no date, no next step and nothing firm about it.

GPT-5.5 (71 words) asked the same question, then added “If there's an issue with the invoice, let me know by the end of the day so we can resolve it promptly.” The deadline is for raising problems rather than for paying, but it at least gives Megan something to act on, and it left a name placeholder at the end.

Both found all three bugs in a three-line function meant to return the second-largest number: it changes the caller's list, crashes on short lists and mishandles duplicates. Both rewrote it to raise a clear error. The difference was in how they treated duplicates. Grok called it a plain bug: “Fails with duplicates (e.g. [1,2,2] wrongly returns 2)” GPT-5.5 said the function “returns a duplicate maximum ([3, 3, 2] -> 3) if "second-largest" means second-largest distinct value”, which is the more accurate way to put it. Some callers want 2 there. That is a narrow win, and Grok's shorter fix is arguably easier to read.

On compound interest both reached Rs 12,597.12; Grok did it in 81 words and an image a teenager would remember: “so your money grows faster, like a snowball rolling downhill.” On the memo, Grok left out the 94% starting figure for fast replies but asked the better question: “What metrics should be tracked if the pilot runs during a busier period?” GPT-5.5 kept every number and asked a question the memo had already raised. On the jug puzzle (4 litres from a 6 and a 9-litre jug, which can't be done) both said no and gave the reason.

Grok 4.3 is a reasoning model, and on these prompts it reasoned at length: 2,808 of the 3,462 output tokens it was billed for were thinking you never see, against 186 for GPT-5.5. Even so, its output price is so much lower ($2.50 per million tokens against $30.00) that the whole answer still cost less. One more oddity: every Grok call was billed for exactly 186 more input tokens than GPT-5.5 counted for the same prompt. We sent no system prompt, so these look like a fixed preamble added before ours. At $1.25 per million they cost almost nothing.

Those are API prices, which is what you pay when a developer tool or an app calls the model for you. The ChatGPT app sells monthly plans instead, listed on AI subscription prices; xAI's SuperGrok price wasn't on any page we could read when we checked. The Claude and Gemini results are on ChatGPT vs Claude and Gemini vs ChatGPT.

Grok 4.3

Speed first: 33 to 47 seconds per answer is fine for a batch job and painful in a chat window. Its writing does the minimum, so an email that needed a deadline came back without one, and a summary lost a baseline figure. It also states its own reading of a vague request as fact, as it did with the duplicates in the code task.

GPT-5.5

It broke the one hard number in the trip prompt, noticed, and patched it with a sentence instead of fixing the plan. Its memo question handed the finance lead's point straight back instead of adding one. And per answer it cost 3.9 times as much as Grok on the same prompts.

Neither one is a substitute for a professional on legal, medical or tax questions. Both will answer confidently, and in our test both made at least one confident claim (Grok on duplicates, GPT-5.5 on its budget) that needed a second look.

Grok has moved on quickly. OpenRouter lists Grok 4.7 (released 21 Sep 2026, $2.00 in and $6.00 out per million tokens), and OpenAI's GPT-6.1 Terra (26 Oct 2026, $1.00 and $5.00). We haven't run our prompts through either, so this page says nothing about how they compare.